M2 delegated 40% of marketing texts to AI and maintained content quality

M2 has demonstrated a practical scenario in which AI does not replace the editorial team but takes on a significant portion of routine content. Within the company, neural networks have already been entrusted with approximately 40% of product and marketing texts, but the final scheme is not reduced to pressing a single button: quality is maintained through strict rules, precise prompts, and mandatory human involvement in complex and creative tasks. The reason for launching the AI editorial was a growth in content volume.

The editorial team needed to cover more tasks for marketing, sales, and support without expanding staff or outsourcing everything. At first, M2 considered standard options — hiring a new copywriter, outsourcing part of the work, or building an internal AI tool. They chose the third path as the fastest and most economical, but immediately imposed a limitation: the neural network must write in the brand's tone and comply with internal editorial policy no worse than a junior author.

The key problem turned out not to be in the choice of model, but in how to explain the company's standards to it. M2 did not train a separate model on its own text corpus but went through a prompt system. At the first stage, the team collected the main Tone of Voice rules, syntactic constraints, stop words, and product naming requirements into one large system prompt.

This zero-shot approach did not work: the model lost focus, wrote too flatly, and sometimes mixed styles. After that, the editorial team moved to a few-shot scenario, where good and bad examples of real texts were added to the prompt along with instructions. This gave predictable results and made it possible to create separate templates for news, social media, and other formats.

Technically, the AI editorial is built on the corporate Videocat platform with llama.cpp and Open WebUI components, and Google Gemma 3 is used as the working model in the described case. But in this story, what matters more than the specific stack is the architectural principle: the tool remains flexible so that different models can be connected and tested for different tasks.

This approach removes dependence on a single provider and allows you to find the right balance between quality, speed, and generation cost. As a result, the neural network at M2 not only writes drafts and adapts texts for channels but also helps generate ideas, transcribe information from images, and explain its own logic. The latter is especially important for employees outside the editorial team: when the model argues for the choice of words and structure, it simultaneously works as an assistant and as a training tool.

At the same time, the company deliberately does not try to turn AI into a full replacement for an editor. M2 clearly divides areas of responsibility. Routine, template-based, and mass tasks can be given to the machine almost entirely, but in-depth texts, creative concepts, slogans, and materials where emotional precision is needed still remain with people.

This balance helps relieve the anxiety of a blank page, speed up the preparation of drafts, and relieve the editorial team without losing the humanity of the text. A separate section is security: the AI tool operates within an internal circuit, and employees are trained on what data can be uploaded to the system and what cannot. This is an important detail for any company that wants to use generative models not in a sandbox but in real processes.

The conclusion from the M2 case is quite down-to-earth and therefore useful: neural networks can already take a significant portion of content routine, but they cannot by themselves create strong editorial work. First, companies need their own standards, a clear Tone of Voice, and live editors who distinguish between good text and formally correct text. Only after that does AI become not a threat to jobs but a force multiplier for the team.

For the market, this is another signal: demand will shift not from authors to machines, but from ordinary performers to specialists who can design processes, set up prompts, and bring machine drafts to the level of a brand.